基于改进YOLOv5s的铸坯表面缺陷检测系统  

Casting Blank Surface Defects Detection System Based on Improved YOLOv5s

作　　者：邓能辉 周秉国 张志杰 石杰[1,2] 吴昆鹏 DENG Nenghui;ZHOU Bingguo;ZHANG Zhijie;SHI Jie;WU Kunpeng(National Engineering Technology Research Center of Flat Rolling Equipment,University of Science and Technology Beijing,Beijing 100083,China;Design and Research Institute of USTB Co.,Ltd.,Beijing 100083,China;Lu'an Iron and Steel Holding Group Co.,Ltd,Lu'an 237400,China)

机构地区：[1]北京科技大学国家板带生产先进装备工程技术研究中心,北京100083 [2]北京科技大学设计研究院有限公司,北京100083 [3]六安钢铁控股集团有限公司,安徽六安237400

出　　处：《仪表技术与传感器》2023年第10期72-78,共7页Instrument Technique and Sensor

基　　金：国家自然科学基金(52004029);广西科技重大专项(AA22068080)。

摘　　要：针对目前连铸坯表面缺陷检测方法存在检测准确率和效率低的问题,提出了一种基于改进YOLOv5s的连铸坯表面缺陷检测系统。首先,基于CycleGAN的域迁移能力和冷轧样本集实现铸坯复杂背景的简单化。其次,利用Ghost网络和GhostBottleneck重新构建YOLOv5s的特征提取骨架以达到轻量化网络结构提高检测速度的目的。最后,在YOLOv5s颈部模块中嵌入SE注意力机制以提升缺陷关键信息捕捉能力,从而提高检测准确率。实验结果表明,改进YOLOv5s在铸坯表面图像数据集上mAP指标达到93.6%,相较于原始的YOLOv5s,mAP指标提升了2.9%,计算量降低了2.5 FLOPs。能够满足铸坯表面缺陷检测系统的实时要求及准确率指标,并且降低了部署所需的计算资源。In response to the problems of low detection accuracy and efficiency in current surface defects detection methods for casting blank,this paper proposed a surface defects detection system for casting blank based on improved YOLOv5s.Firstly,based on the domain adaptation ability of CycleGAN and the cold rolled sample dataset,the complex background of the casting blank was simplified.Secondly,the feature extraction skeleton of YOLOv5s was reconstructed using Ghost network and GhostBottleneck to achieve the goal of lightweight network structure and improved detection speed.Finally,the SE attention mechanism was embedded in the YOLOv5s neck module to improve the ability to capture critical information of defects,thereby improving detection accuracy.The experimental results show that the mAP of the improved YOLOv5s on the surface image dataset of the casting blank reaches 93.6%,which increases by 2.9%compared to the original YOLOv5s and the computational complexity is reduced by 2.5 FLOPs.The improved YOLOv5s can meet the real-time and accuracy requirements of the casting blank surface defects detection system,and computational resources required for deployment are reduced.

关 键 词：铸坯 缺陷检测 CycleGAN YOLOv5s GHOST GhostBottleneck SE注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]